ACGP: Adaptable Constrained Genetic Programming

Abstract

Genetic Programming requires that allfunctions/terminals (tree labels) be given a priori. Inthe absence of specific information about the solution,the user is often forced to provide a large set, thusenlarging the search space often resulting in reducingthe search efficiency. Moreover, based on heuristics,syntactic constraints, or data typing, a given subtreemay be undesired or invalid in a given context. TypedGenetic Programming methods give users the power tospecify some rules for valid tree construction, andthus to prune the otherwise unconstrainedrepresentation in which Genetic Programming operates.However, in general, the user may not be aware of thebest representation space to solve a particularproblem. Moreover, some information may be in the formof weak heuristics. In this work, we present amethodology, which automatically adapts therepresentation for solving a particular problem, byextracting and using such heuristics. Even though manyspecific techniques can be implemented in themethodology, in this paper we use information on localfirst-order (parent-child) distributions of thefunctions and terminals. The heuristics are extractedfrom the population by observing their distribution inbetter individuals. The methodology is illustrated andvalidated using a number of experiments with the11-multiplexer. Moreover, some preliminary empiricalresults linking population size and the sampling rateare also given.